Update:

Unload the BERT model.
Ignore loading warning.

ModelManager.py

@@ -253,6 +253,10 @@ def unload_model(self, model_type_value: str, model_id: str):
                     if key == model_id:
                         break
                     start += ns
+
+                if model_type == ModelType.BERT_VITS2:
+                    for bert_model_name in self.models[model_type][model_id][1].bert_model_names.values():
+                        self.bert_handler.release_bert(bert_model_name)
 
                 del self.sid2model[model_type][start:start + n_speakers]
                 del self.voice_speakers[model_type.value][start:start + n_speakers]
@@ -266,6 +270,7 @@ def unload_model(self, model_type_value: str, model_id: str):
 
                 state = True
                 self.notify("model_unloaded", model_manager=self)
+                self.logger.info(f"Unloading success.")
         except Exception as e:
             self.logger.info(f"Unloading failed. {e}")
             state = False

bert_vits2/text/bert_handler.py

@@ -1,10 +1,11 @@
+import gc
 import logging
 import os
 
+import torch
 from transformers import AutoTokenizer, AutoModelForMaskedLM
 
 from utils.config_manager import global_config as config
-from logger import logger
 from utils.download import download_file
 from .chinese_bert import get_bert_feature as zh_bert
 from .english_bert_mock import get_bert_feature as en_bert
@@ -53,7 +54,7 @@ def _download_model(self, bert_model_name, target_path=None):
                 "https://hf-mirror.com/microsoft/deberta-v3-large/resolve/main/spm.model",
             ],
         }
-        
+
         SHA256 = {
             "CHINESE_ROBERTA_WWM_EXT_LARGE": "4ac62d49144d770c5ca9a5d1d3039c4995665a080febe63198189857c6bd11cd",
             "BERT_BASE_JAPANESE_V3": "e172862e0674054d65e0ba40d67df2a4687982f589db44aa27091c386e5450a4",
@@ -70,9 +71,9 @@ def _download_model(self, bert_model_name, target_path=None):
         expected_sha256 = SHA256[bert_model_name]
         success, message = download_file(urls, target_path, expected_sha256=expected_sha256)
         if not success:
-            logger.error(f"Failed to download {bert_model_name}: {message}")
+            logging.error(f"Failed to download {bert_model_name}: {message}")
         else:
-            logger.info(f"{message}")
+            logging.info(f"{message}")
 
     def load_bert(self, bert_model_name, max_retries=3):
         if bert_model_name not in self.bert_models:
@@ -119,7 +120,9 @@ def release_bert(self, bert_model_name):
             if count == 0:
                 # 当引用计数为0时，删除模型并释放其资源
                 del self.bert_models[bert_model_name]
-                logger(f"Model {bert_model_name} has been released.")
+                gc.collect()
+                torch.cuda.empty_cache()
+                logging.info(f"BERT model {bert_model_name} has been released.")
             else:
                 tokenizer, model = self.bert_models[bert_model_name][:2]
                 self.bert_models[bert_model_name] = (tokenizer, model, count)

bert_vits2/utils.py

@@ -47,6 +47,12 @@ def load_checkpoint(checkpoint_path, model, optimizer=None, skip_optimizer=False
                 logger.error(f"{k} is not in the checkpoint")
                 logger.warning(
                     f"If you're using a transitional version, please add the \"version\": \"1.1.0-transition\" parameter within the \"data\" section of the model's config.json.")
+            elif "en_bert_proj" in k:
+                v = torch.zeros_like(v)
+                if version is None:
+                    logger.error(f"{k} is not in the checkpoint")
+                    logger.warning(
+                        f"If you're using an older version of the model, consider adding the \"version\" parameter to the model's config.json under the \"data\" section. For instance: \"legacy_version\": \"1.1.1\"")
             else:
                 logger.error(f"{k} is not in the checkpoint")